In-training Restoration Models Matter: Data Augmentation for Degraded-reference Image Quality Assessment

Abstract

Full-Reference Image Quality Assessment (FR-IQA) metrics such as PSNR, SSIM, and LPIPS have been widely adopted for evaluating image restoration (IR) methods. However, pristine-quality images are usually not available, making inferior No-Reference Image Quality Assessment (NR-IQA) metrics seem to be the only solutions in practical applications. Fortunately, when evaluating image restoration methods, paired degraded and restoration images are generally available. Thus, this paper takes a step forward to develop a Degraded-Reference IQA (DR-IQA) model while respecting its correspondence with FR-IQA metrics. To this end, we adopt a simple encoder-decoder as DR-IQA model, and take paired degraded and restoration images as the input to predict distortion maps guided by FR-IQA metrics. More importantly, due to the diversity and continuous development of image restoration models, it is difficult to make the DR-IQA model learned based on a specific restoration model generalize well to other ones. To address this issue, we augment the DR-IQA training samples by adding the results produced by in-training restoration models. Benefiting from the diversity of training samples, our learned DR-IQA model generalizes well to unseen restoration models. We respectively test our DR-IQA models on various image restoration tasks,e.g., denoising, super-resolution, JPEG deblocking, and complicated degradations, where our method can further close the performance gap between FR-IQA metrics and the state-of-the-art NR-IQA methods. Moreover, experiments also show the effectiveness of our method in performance comparison and model selection of image restoration models without ground-truth clean images. Source code will be made publicly available.